I build AI applications and run model experiments. My work spans OCR research, small-model training and sandbox-based evaluation. My technical focus includes reinforcement learning and post-training—training objectives, reward design and how we measure whether a model actually improves.
Selected Work
Built seed-matched, single-GPU experiments comparing AdamW and Hybrid Muon on small GPT models. The public report shows validation-loss and throughput tradeoffs, with the benchmark's limits stated.
Built a proof of concept to detect cancelled handwriting before OCR. The repository documents detector comparisons, confidence calibration and cases where the system should abstain. It is not deployed.
Built a bounded tool loop that runs Python tasks in Docker. The public repository shows a prototype and its execution flow; it does not claim a security audit or production scale.
Supporting Projects
Built a multimodal AI workspace with model routing, visual learning outputs and a multi-step research mode.
Built an S3 desktop browser with resumable SSH/SFTP transfers and tests for interruption and restart recovery.
Additional Work
Full-stack application with Solidity contracts acting as the system of record. Contract logic is tested with Foundry.
Modular protocol core using the Diamond Standard — facets for access control and execution logic, strict storage layout conventions to prevent slot collisions across upgrades.
Smart contracts reasoning about state and interactions across L3 and L2 layers. Focused on cross-layer assumptions, state consistency, and correct execution semantics.
Languages
Python, TypeScript, JavaScript, Solidity, SQL
AI & Machine Learning
Model Training, Model Evaluation, Reinforcement Learning, Post-Training, Computer Vision, OCR, PyTorch
Frontend
React, React Native, Next.js, Tailwind CSS
Backend
Node.js, Express, REST APIs, PostgreSQL
Cloud & Infra
Docker, AWS (S3, RDS, Fargate)
Web3
Solidity, EVM, Foundry, Smart Contracts
AI applications, model experiments, OCR research, evaluation and agent systems. His technical focus also includes reinforcement learning and post-training.
Reinforcement learning and post-training are technical focus areas. The portfolio links to model training and evaluation work; it does not claim a completed RL system.
Python, PyTorch, TypeScript, Node.js, Docker, React, PostgreSQL and AWS, as shown across the linked projects.
Email anubrat23@gmail.com or connect on LinkedIn at linkedin.com/in/anubrat-sahoo.